2 * Note: this file was generated by the Gromacs avx_256_double kernel generator.
4 * This source code is part of
8 * Copyright (c) 2001-2012, The GROMACS Development Team
10 * Gromacs is a library for molecular simulation and trajectory analysis,
11 * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
12 * a full list of developers and information, check out http://www.gromacs.org
14 * This program is free software; you can redistribute it and/or modify it under
15 * the terms of the GNU Lesser General Public License as published by the Free
16 * Software Foundation; either version 2 of the License, or (at your option) any
19 * To help fund GROMACS development, we humbly ask that you cite
20 * the papers people have written on it - you can find them on the website.
28 #include "../nb_kernel.h"
29 #include "types/simple.h"
33 #include "gmx_math_x86_avx_256_double.h"
34 #include "kernelutil_x86_avx_256_double.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_256_double
38 * Electrostatics interaction: ReactionField
39 * VdW interaction: LennardJones
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_256_double
45 (t_nblist * gmx_restrict nlist,
46 rvec * gmx_restrict xx,
47 rvec * gmx_restrict ff,
48 t_forcerec * gmx_restrict fr,
49 t_mdatoms * gmx_restrict mdatoms,
50 nb_kernel_data_t * gmx_restrict kernel_data,
51 t_nrnb * gmx_restrict nrnb)
53 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
54 * just 0 for non-waters.
55 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
56 * jnr indices corresponding to data put in the four positions in the SIMD register.
58 int i_shift_offset,i_coord_offset,outeriter,inneriter;
59 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
60 int jnrA,jnrB,jnrC,jnrD;
61 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
62 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
63 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
64 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
66 real *shiftvec,*fshift,*x,*f;
67 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
69 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
70 real * vdwioffsetptr0;
71 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
72 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
73 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
74 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
75 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
78 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
81 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
82 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
83 __m256d rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
84 real rswitch_scalar,d_scalar;
85 __m256d dummy_mask,cutoff_mask;
86 __m128 tmpmask0,tmpmask1;
87 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
88 __m256d one = _mm256_set1_pd(1.0);
89 __m256d two = _mm256_set1_pd(2.0);
95 jindex = nlist->jindex;
97 shiftidx = nlist->shift;
99 shiftvec = fr->shift_vec[0];
100 fshift = fr->fshift[0];
101 facel = _mm256_set1_pd(fr->epsfac);
102 charge = mdatoms->chargeA;
103 krf = _mm256_set1_pd(fr->ic->k_rf);
104 krf2 = _mm256_set1_pd(fr->ic->k_rf*2.0);
105 crf = _mm256_set1_pd(fr->ic->c_rf);
106 nvdwtype = fr->ntype;
108 vdwtype = mdatoms->typeA;
110 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
111 rcutoff_scalar = fr->rcoulomb;
112 rcutoff = _mm256_set1_pd(rcutoff_scalar);
113 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
115 rswitch_scalar = fr->rvdw_switch;
116 rswitch = _mm256_set1_pd(rswitch_scalar);
117 /* Setup switch parameters */
118 d_scalar = rcutoff_scalar-rswitch_scalar;
119 d = _mm256_set1_pd(d_scalar);
120 swV3 = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
121 swV4 = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
122 swV5 = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
123 swF2 = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
124 swF3 = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
125 swF4 = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
127 /* Avoid stupid compiler warnings */
128 jnrA = jnrB = jnrC = jnrD = 0;
137 for(iidx=0;iidx<4*DIM;iidx++)
142 /* Start outer loop over neighborlists */
143 for(iidx=0; iidx<nri; iidx++)
145 /* Load shift vector for this list */
146 i_shift_offset = DIM*shiftidx[iidx];
148 /* Load limits for loop over neighbors */
149 j_index_start = jindex[iidx];
150 j_index_end = jindex[iidx+1];
152 /* Get outer coordinate index */
154 i_coord_offset = DIM*inr;
156 /* Load i particle coords and add shift vector */
157 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
159 fix0 = _mm256_setzero_pd();
160 fiy0 = _mm256_setzero_pd();
161 fiz0 = _mm256_setzero_pd();
163 /* Load parameters for i particles */
164 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
165 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
167 /* Reset potential sums */
168 velecsum = _mm256_setzero_pd();
169 vvdwsum = _mm256_setzero_pd();
171 /* Start inner kernel loop */
172 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
175 /* Get j neighbor index, and coordinate index */
180 j_coord_offsetA = DIM*jnrA;
181 j_coord_offsetB = DIM*jnrB;
182 j_coord_offsetC = DIM*jnrC;
183 j_coord_offsetD = DIM*jnrD;
185 /* load j atom coordinates */
186 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
187 x+j_coord_offsetC,x+j_coord_offsetD,
190 /* Calculate displacement vector */
191 dx00 = _mm256_sub_pd(ix0,jx0);
192 dy00 = _mm256_sub_pd(iy0,jy0);
193 dz00 = _mm256_sub_pd(iz0,jz0);
195 /* Calculate squared distance and things based on it */
196 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
198 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
200 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
202 /* Load parameters for j particles */
203 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
204 charge+jnrC+0,charge+jnrD+0);
205 vdwjidx0A = 2*vdwtype[jnrA+0];
206 vdwjidx0B = 2*vdwtype[jnrB+0];
207 vdwjidx0C = 2*vdwtype[jnrC+0];
208 vdwjidx0D = 2*vdwtype[jnrD+0];
210 /**************************
211 * CALCULATE INTERACTIONS *
212 **************************/
214 if (gmx_mm256_any_lt(rsq00,rcutoff2))
217 r00 = _mm256_mul_pd(rsq00,rinv00);
219 /* Compute parameters for interactions between i and j atoms */
220 qq00 = _mm256_mul_pd(iq0,jq0);
221 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
222 vdwioffsetptr0+vdwjidx0B,
223 vdwioffsetptr0+vdwjidx0C,
224 vdwioffsetptr0+vdwjidx0D,
227 /* REACTION-FIELD ELECTROSTATICS */
228 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
229 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
231 /* LENNARD-JONES DISPERSION/REPULSION */
233 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
234 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
235 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
236 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
237 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
239 d = _mm256_sub_pd(r00,rswitch);
240 d = _mm256_max_pd(d,_mm256_setzero_pd());
241 d2 = _mm256_mul_pd(d,d);
242 sw = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
244 dsw = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
246 /* Evaluate switch function */
247 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
248 fvdw = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
249 vvdw = _mm256_mul_pd(vvdw,sw);
250 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
252 /* Update potential sum for this i atom from the interaction with this j atom. */
253 velec = _mm256_and_pd(velec,cutoff_mask);
254 velecsum = _mm256_add_pd(velecsum,velec);
255 vvdw = _mm256_and_pd(vvdw,cutoff_mask);
256 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
258 fscal = _mm256_add_pd(felec,fvdw);
260 fscal = _mm256_and_pd(fscal,cutoff_mask);
262 /* Calculate temporary vectorial force */
263 tx = _mm256_mul_pd(fscal,dx00);
264 ty = _mm256_mul_pd(fscal,dy00);
265 tz = _mm256_mul_pd(fscal,dz00);
267 /* Update vectorial force */
268 fix0 = _mm256_add_pd(fix0,tx);
269 fiy0 = _mm256_add_pd(fiy0,ty);
270 fiz0 = _mm256_add_pd(fiz0,tz);
272 fjptrA = f+j_coord_offsetA;
273 fjptrB = f+j_coord_offsetB;
274 fjptrC = f+j_coord_offsetC;
275 fjptrD = f+j_coord_offsetD;
276 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
280 /* Inner loop uses 70 flops */
286 /* Get j neighbor index, and coordinate index */
287 jnrlistA = jjnr[jidx];
288 jnrlistB = jjnr[jidx+1];
289 jnrlistC = jjnr[jidx+2];
290 jnrlistD = jjnr[jidx+3];
291 /* Sign of each element will be negative for non-real atoms.
292 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
293 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
295 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
297 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
298 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
299 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
301 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
302 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
303 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
304 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
305 j_coord_offsetA = DIM*jnrA;
306 j_coord_offsetB = DIM*jnrB;
307 j_coord_offsetC = DIM*jnrC;
308 j_coord_offsetD = DIM*jnrD;
310 /* load j atom coordinates */
311 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
312 x+j_coord_offsetC,x+j_coord_offsetD,
315 /* Calculate displacement vector */
316 dx00 = _mm256_sub_pd(ix0,jx0);
317 dy00 = _mm256_sub_pd(iy0,jy0);
318 dz00 = _mm256_sub_pd(iz0,jz0);
320 /* Calculate squared distance and things based on it */
321 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
323 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
325 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
327 /* Load parameters for j particles */
328 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
329 charge+jnrC+0,charge+jnrD+0);
330 vdwjidx0A = 2*vdwtype[jnrA+0];
331 vdwjidx0B = 2*vdwtype[jnrB+0];
332 vdwjidx0C = 2*vdwtype[jnrC+0];
333 vdwjidx0D = 2*vdwtype[jnrD+0];
335 /**************************
336 * CALCULATE INTERACTIONS *
337 **************************/
339 if (gmx_mm256_any_lt(rsq00,rcutoff2))
342 r00 = _mm256_mul_pd(rsq00,rinv00);
343 r00 = _mm256_andnot_pd(dummy_mask,r00);
345 /* Compute parameters for interactions between i and j atoms */
346 qq00 = _mm256_mul_pd(iq0,jq0);
347 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
348 vdwioffsetptr0+vdwjidx0B,
349 vdwioffsetptr0+vdwjidx0C,
350 vdwioffsetptr0+vdwjidx0D,
353 /* REACTION-FIELD ELECTROSTATICS */
354 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
355 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
357 /* LENNARD-JONES DISPERSION/REPULSION */
359 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
360 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
361 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
362 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
363 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
365 d = _mm256_sub_pd(r00,rswitch);
366 d = _mm256_max_pd(d,_mm256_setzero_pd());
367 d2 = _mm256_mul_pd(d,d);
368 sw = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
370 dsw = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
372 /* Evaluate switch function */
373 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
374 fvdw = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
375 vvdw = _mm256_mul_pd(vvdw,sw);
376 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
378 /* Update potential sum for this i atom from the interaction with this j atom. */
379 velec = _mm256_and_pd(velec,cutoff_mask);
380 velec = _mm256_andnot_pd(dummy_mask,velec);
381 velecsum = _mm256_add_pd(velecsum,velec);
382 vvdw = _mm256_and_pd(vvdw,cutoff_mask);
383 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
384 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
386 fscal = _mm256_add_pd(felec,fvdw);
388 fscal = _mm256_and_pd(fscal,cutoff_mask);
390 fscal = _mm256_andnot_pd(dummy_mask,fscal);
392 /* Calculate temporary vectorial force */
393 tx = _mm256_mul_pd(fscal,dx00);
394 ty = _mm256_mul_pd(fscal,dy00);
395 tz = _mm256_mul_pd(fscal,dz00);
397 /* Update vectorial force */
398 fix0 = _mm256_add_pd(fix0,tx);
399 fiy0 = _mm256_add_pd(fiy0,ty);
400 fiz0 = _mm256_add_pd(fiz0,tz);
402 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
403 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
404 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
405 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
406 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
410 /* Inner loop uses 71 flops */
413 /* End of innermost loop */
415 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
416 f+i_coord_offset,fshift+i_shift_offset);
419 /* Update potential energies */
420 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
421 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
423 /* Increment number of inner iterations */
424 inneriter += j_index_end - j_index_start;
426 /* Outer loop uses 9 flops */
429 /* Increment number of outer iterations */
432 /* Update outer/inner flops */
434 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*71);
437 * Gromacs nonbonded kernel: nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_256_double
438 * Electrostatics interaction: ReactionField
439 * VdW interaction: LennardJones
440 * Geometry: Particle-Particle
441 * Calculate force/pot: Force
444 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_256_double
445 (t_nblist * gmx_restrict nlist,
446 rvec * gmx_restrict xx,
447 rvec * gmx_restrict ff,
448 t_forcerec * gmx_restrict fr,
449 t_mdatoms * gmx_restrict mdatoms,
450 nb_kernel_data_t * gmx_restrict kernel_data,
451 t_nrnb * gmx_restrict nrnb)
453 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
454 * just 0 for non-waters.
455 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
456 * jnr indices corresponding to data put in the four positions in the SIMD register.
458 int i_shift_offset,i_coord_offset,outeriter,inneriter;
459 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
460 int jnrA,jnrB,jnrC,jnrD;
461 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
462 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
463 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
464 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
466 real *shiftvec,*fshift,*x,*f;
467 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
469 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
470 real * vdwioffsetptr0;
471 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
472 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
473 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
474 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
475 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
478 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
481 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
482 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
483 __m256d rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
484 real rswitch_scalar,d_scalar;
485 __m256d dummy_mask,cutoff_mask;
486 __m128 tmpmask0,tmpmask1;
487 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
488 __m256d one = _mm256_set1_pd(1.0);
489 __m256d two = _mm256_set1_pd(2.0);
495 jindex = nlist->jindex;
497 shiftidx = nlist->shift;
499 shiftvec = fr->shift_vec[0];
500 fshift = fr->fshift[0];
501 facel = _mm256_set1_pd(fr->epsfac);
502 charge = mdatoms->chargeA;
503 krf = _mm256_set1_pd(fr->ic->k_rf);
504 krf2 = _mm256_set1_pd(fr->ic->k_rf*2.0);
505 crf = _mm256_set1_pd(fr->ic->c_rf);
506 nvdwtype = fr->ntype;
508 vdwtype = mdatoms->typeA;
510 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
511 rcutoff_scalar = fr->rcoulomb;
512 rcutoff = _mm256_set1_pd(rcutoff_scalar);
513 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
515 rswitch_scalar = fr->rvdw_switch;
516 rswitch = _mm256_set1_pd(rswitch_scalar);
517 /* Setup switch parameters */
518 d_scalar = rcutoff_scalar-rswitch_scalar;
519 d = _mm256_set1_pd(d_scalar);
520 swV3 = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
521 swV4 = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
522 swV5 = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
523 swF2 = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
524 swF3 = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
525 swF4 = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
527 /* Avoid stupid compiler warnings */
528 jnrA = jnrB = jnrC = jnrD = 0;
537 for(iidx=0;iidx<4*DIM;iidx++)
542 /* Start outer loop over neighborlists */
543 for(iidx=0; iidx<nri; iidx++)
545 /* Load shift vector for this list */
546 i_shift_offset = DIM*shiftidx[iidx];
548 /* Load limits for loop over neighbors */
549 j_index_start = jindex[iidx];
550 j_index_end = jindex[iidx+1];
552 /* Get outer coordinate index */
554 i_coord_offset = DIM*inr;
556 /* Load i particle coords and add shift vector */
557 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
559 fix0 = _mm256_setzero_pd();
560 fiy0 = _mm256_setzero_pd();
561 fiz0 = _mm256_setzero_pd();
563 /* Load parameters for i particles */
564 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
565 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
567 /* Start inner kernel loop */
568 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
571 /* Get j neighbor index, and coordinate index */
576 j_coord_offsetA = DIM*jnrA;
577 j_coord_offsetB = DIM*jnrB;
578 j_coord_offsetC = DIM*jnrC;
579 j_coord_offsetD = DIM*jnrD;
581 /* load j atom coordinates */
582 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
583 x+j_coord_offsetC,x+j_coord_offsetD,
586 /* Calculate displacement vector */
587 dx00 = _mm256_sub_pd(ix0,jx0);
588 dy00 = _mm256_sub_pd(iy0,jy0);
589 dz00 = _mm256_sub_pd(iz0,jz0);
591 /* Calculate squared distance and things based on it */
592 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
594 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
596 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
598 /* Load parameters for j particles */
599 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
600 charge+jnrC+0,charge+jnrD+0);
601 vdwjidx0A = 2*vdwtype[jnrA+0];
602 vdwjidx0B = 2*vdwtype[jnrB+0];
603 vdwjidx0C = 2*vdwtype[jnrC+0];
604 vdwjidx0D = 2*vdwtype[jnrD+0];
606 /**************************
607 * CALCULATE INTERACTIONS *
608 **************************/
610 if (gmx_mm256_any_lt(rsq00,rcutoff2))
613 r00 = _mm256_mul_pd(rsq00,rinv00);
615 /* Compute parameters for interactions between i and j atoms */
616 qq00 = _mm256_mul_pd(iq0,jq0);
617 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
618 vdwioffsetptr0+vdwjidx0B,
619 vdwioffsetptr0+vdwjidx0C,
620 vdwioffsetptr0+vdwjidx0D,
623 /* REACTION-FIELD ELECTROSTATICS */
624 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
626 /* LENNARD-JONES DISPERSION/REPULSION */
628 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
629 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
630 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
631 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
632 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
634 d = _mm256_sub_pd(r00,rswitch);
635 d = _mm256_max_pd(d,_mm256_setzero_pd());
636 d2 = _mm256_mul_pd(d,d);
637 sw = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
639 dsw = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
641 /* Evaluate switch function */
642 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
643 fvdw = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
644 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
646 fscal = _mm256_add_pd(felec,fvdw);
648 fscal = _mm256_and_pd(fscal,cutoff_mask);
650 /* Calculate temporary vectorial force */
651 tx = _mm256_mul_pd(fscal,dx00);
652 ty = _mm256_mul_pd(fscal,dy00);
653 tz = _mm256_mul_pd(fscal,dz00);
655 /* Update vectorial force */
656 fix0 = _mm256_add_pd(fix0,tx);
657 fiy0 = _mm256_add_pd(fiy0,ty);
658 fiz0 = _mm256_add_pd(fiz0,tz);
660 fjptrA = f+j_coord_offsetA;
661 fjptrB = f+j_coord_offsetB;
662 fjptrC = f+j_coord_offsetC;
663 fjptrD = f+j_coord_offsetD;
664 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
668 /* Inner loop uses 61 flops */
674 /* Get j neighbor index, and coordinate index */
675 jnrlistA = jjnr[jidx];
676 jnrlistB = jjnr[jidx+1];
677 jnrlistC = jjnr[jidx+2];
678 jnrlistD = jjnr[jidx+3];
679 /* Sign of each element will be negative for non-real atoms.
680 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
681 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
683 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
685 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
686 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
687 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
689 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
690 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
691 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
692 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
693 j_coord_offsetA = DIM*jnrA;
694 j_coord_offsetB = DIM*jnrB;
695 j_coord_offsetC = DIM*jnrC;
696 j_coord_offsetD = DIM*jnrD;
698 /* load j atom coordinates */
699 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
700 x+j_coord_offsetC,x+j_coord_offsetD,
703 /* Calculate displacement vector */
704 dx00 = _mm256_sub_pd(ix0,jx0);
705 dy00 = _mm256_sub_pd(iy0,jy0);
706 dz00 = _mm256_sub_pd(iz0,jz0);
708 /* Calculate squared distance and things based on it */
709 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
711 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
713 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
715 /* Load parameters for j particles */
716 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
717 charge+jnrC+0,charge+jnrD+0);
718 vdwjidx0A = 2*vdwtype[jnrA+0];
719 vdwjidx0B = 2*vdwtype[jnrB+0];
720 vdwjidx0C = 2*vdwtype[jnrC+0];
721 vdwjidx0D = 2*vdwtype[jnrD+0];
723 /**************************
724 * CALCULATE INTERACTIONS *
725 **************************/
727 if (gmx_mm256_any_lt(rsq00,rcutoff2))
730 r00 = _mm256_mul_pd(rsq00,rinv00);
731 r00 = _mm256_andnot_pd(dummy_mask,r00);
733 /* Compute parameters for interactions between i and j atoms */
734 qq00 = _mm256_mul_pd(iq0,jq0);
735 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
736 vdwioffsetptr0+vdwjidx0B,
737 vdwioffsetptr0+vdwjidx0C,
738 vdwioffsetptr0+vdwjidx0D,
741 /* REACTION-FIELD ELECTROSTATICS */
742 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
744 /* LENNARD-JONES DISPERSION/REPULSION */
746 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
747 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
748 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
749 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
750 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
752 d = _mm256_sub_pd(r00,rswitch);
753 d = _mm256_max_pd(d,_mm256_setzero_pd());
754 d2 = _mm256_mul_pd(d,d);
755 sw = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
757 dsw = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
759 /* Evaluate switch function */
760 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
761 fvdw = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
762 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
764 fscal = _mm256_add_pd(felec,fvdw);
766 fscal = _mm256_and_pd(fscal,cutoff_mask);
768 fscal = _mm256_andnot_pd(dummy_mask,fscal);
770 /* Calculate temporary vectorial force */
771 tx = _mm256_mul_pd(fscal,dx00);
772 ty = _mm256_mul_pd(fscal,dy00);
773 tz = _mm256_mul_pd(fscal,dz00);
775 /* Update vectorial force */
776 fix0 = _mm256_add_pd(fix0,tx);
777 fiy0 = _mm256_add_pd(fiy0,ty);
778 fiz0 = _mm256_add_pd(fiz0,tz);
780 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
781 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
782 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
783 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
784 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
788 /* Inner loop uses 62 flops */
791 /* End of innermost loop */
793 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
794 f+i_coord_offset,fshift+i_shift_offset);
796 /* Increment number of inner iterations */
797 inneriter += j_index_end - j_index_start;
799 /* Outer loop uses 7 flops */
802 /* Increment number of outer iterations */
805 /* Update outer/inner flops */
807 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*62);